[Chore] Rename `Processor` to `InputProcessor` (#29682)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>

tests/entrypoints/openai/test_lora_resolvers.py

@@ -114,7 +114,7 @@ def mock_serving_setup():
     mock_engine.add_lora.reset_mock()
 
     mock_engine.model_config = MockModelConfig()
-    mock_engine.processor = MagicMock()
+    mock_engine.input_processor = MagicMock()
     mock_engine.io_processor = MagicMock()
 
     models = OpenAIServingModels(

tests/entrypoints/openai/test_serving_chat.py

@@ -429,7 +429,7 @@ async def test_serving_chat_returns_correct_model_name():
     mock_engine.get_tokenizer.return_value = get_tokenizer(MODEL_NAME)
     mock_engine.errored = False
     mock_engine.model_config = MockModelConfig()
-    mock_engine.processor = MagicMock()
+    mock_engine.input_processor = MagicMock()
     mock_engine.io_processor = MagicMock()
 
     serving_chat = _build_serving_chat(mock_engine)
@@ -459,7 +459,7 @@ async def test_serving_chat_should_set_correct_max_tokens():
     mock_engine.get_tokenizer.return_value = get_tokenizer(MODEL_NAME)
     mock_engine.errored = False
     mock_engine.model_config = MockModelConfig()
-    mock_engine.processor = MagicMock()
+    mock_engine.input_processor = MagicMock()
     mock_engine.io_processor = MagicMock()
 
     serving_chat = _build_serving_chat(mock_engine)
@@ -492,7 +492,7 @@ async def test_serving_chat_should_set_correct_max_tokens():
     mock_engine.get_tokenizer.return_value = get_tokenizer(MODEL_NAME)
     mock_engine.errored = False
     mock_engine.model_config = mock_model_config
-    mock_engine.processor = MagicMock()
+    mock_engine.input_processor = MagicMock()
     mock_engine.io_processor = MagicMock()
 
     # Initialize the serving chat
@@ -537,7 +537,7 @@ async def test_serving_chat_should_set_correct_max_tokens():
     mock_engine.get_tokenizer.return_value = get_tokenizer(MODEL_NAME)
     mock_engine.errored = False
     mock_engine.model_config = mock_model_config
-    mock_engine.processor = MagicMock()
+    mock_engine.input_processor = MagicMock()
     mock_engine.io_processor = MagicMock()
 
     # Initialize the serving chat
@@ -583,7 +583,7 @@ async def test_serving_chat_could_load_correct_generation_config():
     mock_engine.get_tokenizer.return_value = get_tokenizer(MODEL_NAME)
     mock_engine.errored = False
     mock_engine.model_config = mock_model_config
-    mock_engine.processor = MagicMock()
+    mock_engine.input_processor = MagicMock()
     mock_engine.io_processor = MagicMock()
 
     # Initialize the serving chat
@@ -629,7 +629,7 @@ async def test_serving_chat_did_set_correct_cache_salt(model_type):
     mock_engine.get_tokenizer.return_value = get_tokenizer(MODEL_NAME)
     mock_engine.errored = False
     mock_engine.model_config = mock_model_config
-    mock_engine.processor = MagicMock()
+    mock_engine.input_processor = MagicMock()
     mock_engine.io_processor = MagicMock()
 
     serving_chat = _build_serving_chat(mock_engine)
@@ -662,7 +662,7 @@ async def test_serving_chat_data_parallel_rank_extraction():
     mock_engine.get_tokenizer.return_value = get_tokenizer(MODEL_NAME)
     mock_engine.errored = False
     mock_engine.model_config = MockModelConfig()
-    mock_engine.processor = MagicMock()
+    mock_engine.input_processor = MagicMock()
     mock_engine.io_processor = MagicMock()
 
     # Mock the generate method to return an async generator

tests/entrypoints/openai/test_serving_engine.py

@@ -23,7 +23,7 @@ def serving() -> OpenAIServing:
     model_config.max_model_len = 32768
     models = Mock(spec=OpenAIServingModels)
     models.model_config = model_config
-    models.processor = Mock()
+    models.input_processor = Mock()
     models.io_processor = Mock()
 
     serving = OpenAIServing(

tests/entrypoints/openai/test_serving_models.py

@@ -30,7 +30,7 @@ async def _async_serving_models_init() -> OpenAIServingModels:
     mock_model_config = MagicMock(spec=ModelConfig)
     mock_model_config.max_model_len = 2048
     mock_engine_client.model_config = mock_model_config
-    mock_engine_client.processor = MagicMock()
+    mock_engine_client.input_processor = MagicMock()
     mock_engine_client.io_processor = MagicMock()
 
     serving_models = OpenAIServingModels(

tests/entrypoints/openai/test_serving_responses.py

@@ -127,7 +127,7 @@ class TestInitializeToolSessions:
         model_config.get_diff_sampling_param.return_value = {}
         engine_client.model_config = model_config
 
-        engine_client.processor = MagicMock()
+        engine_client.input_processor = MagicMock()
         engine_client.io_processor = MagicMock()
 
         models = MagicMock()
@@ -213,7 +213,7 @@ class TestValidateGeneratorInput:
         model_config.get_diff_sampling_param.return_value = {}
         engine_client.model_config = model_config
 
-        engine_client.processor = MagicMock()
+        engine_client.input_processor = MagicMock()
         engine_client.io_processor = MagicMock()
 
         models = MagicMock()

tests/v1/engine/test_processor_multi_modal_uuids.py → tests/v1/engine/test_process_multi_modal_uuids.py

@@ -7,18 +7,17 @@ from vllm.assets.image import ImageAsset
 from vllm.assets.video import VideoAsset
 from vllm.config import CacheConfig, DeviceConfig, ModelConfig, VllmConfig
 from vllm.sampling_params import SamplingParams
-from vllm.v1.engine import processor as processor_mod
-from vllm.v1.engine.processor import Processor
+from vllm.v1.engine import input_processor as input_processor_mod
+from vllm.v1.engine.input_processor import InputProcessor
 
 cherry_pil_image = ImageAsset("cherry_blossom").pil_image
 stop_pil_image = ImageAsset("stop_sign").pil_image
 baby_reading_np_ndarrays = VideoAsset("baby_reading").np_ndarrays
 
 
-# Mock processor for testing
-def _mk_processor(
+def _mock_input_processor(
     monkeypatch, *, mm_cache_gb: float = 4.0, enable_prefix_caching: bool = True
-) -> Processor:
+) -> InputProcessor:
     """
     Create a Processor instance with minimal configuration suitable for unit
     tests without accessing external resources.
@@ -36,7 +35,7 @@ def _mk_processor(
         raising=True,
     )
     monkeypatch.setattr(
-        processor_mod,
+        input_processor_mod,
         "processor_cache_from_config",
         lambda vllm_config, mm_registry: None,
         raising=True,
@@ -65,11 +64,11 @@ def _mk_processor(
         device_config=DeviceConfig(device="cpu"),
     )
 
-    return Processor(vllm_config, tokenizer=None)
+    return InputProcessor(vllm_config, tokenizer=None)
 
 
 def test_multi_modal_uuids_length_mismatch_raises(monkeypatch):
-    processor = _mk_processor(monkeypatch)
+    input_processor = _mock_input_processor(monkeypatch)
 
     prompt = {
         "prompt": "USER: <image>\nDescribe\nASSISTANT:",
@@ -79,7 +78,7 @@ def test_multi_modal_uuids_length_mismatch_raises(monkeypatch):
     }
 
     with pytest.raises(ValueError, match="must have same length as data"):
-        processor.process_inputs(
+        input_processor.process_inputs(
             request_id="req-1",
             prompt=prompt,  # type: ignore[arg-type]
             params=SamplingParams(),
@@ -87,7 +86,7 @@ def test_multi_modal_uuids_length_mismatch_raises(monkeypatch):
 
 
 def test_multi_modal_uuids_missing_modality_raises(monkeypatch):
-    processor = _mk_processor(monkeypatch)
+    input_processor = _mock_input_processor(monkeypatch)
 
     prompt = {
         "prompt": "USER: <image><video>\nDescribe\nASSISTANT:",
@@ -101,7 +100,7 @@ def test_multi_modal_uuids_missing_modality_raises(monkeypatch):
     }
 
     with pytest.raises(ValueError, match="must be provided if multi_modal_data"):
-        processor.process_inputs(
+        input_processor.process_inputs(
             request_id="req-2",
             prompt=prompt,  # type: ignore[arg-type]
             params=SamplingParams(),
@@ -119,7 +118,7 @@ def test_multi_modal_uuids_missing_modality_raises(monkeypatch):
 def test_multi_modal_uuids_accepts_none_and_passes_through(
     monkeypatch, mm_cache_gb: float, enable_prefix_caching: bool
 ):
-    processor = _mk_processor(
+    input_processor = _mock_input_processor(
         monkeypatch,
         mm_cache_gb=mm_cache_gb,
         enable_prefix_caching=enable_prefix_caching,
@@ -137,7 +136,7 @@ def test_multi_modal_uuids_accepts_none_and_passes_through(
 
     # Monkeypatch only the bound preprocess method on this instance
     monkeypatch.setattr(
-        processor.input_preprocessor, "preprocess", fake_preprocess, raising=True
+        input_processor.input_preprocessor, "preprocess", fake_preprocess, raising=True
     )
 
     # Use a consistent two-image scenario across all configurations
@@ -151,7 +150,7 @@ def test_multi_modal_uuids_accepts_none_and_passes_through(
         "multi_modal_uuids": mm_uuids,
     }
 
-    processor.process_inputs(
+    input_processor.process_inputs(
         request_id="req-3",
         prompt=prompt,  # type: ignore[arg-type]
         params=SamplingParams(),
@@ -163,7 +162,9 @@ def test_multi_modal_uuids_accepts_none_and_passes_through(
 def test_multi_modal_uuids_ignored_when_caching_disabled(monkeypatch):
     # When both processor cache is 0 and prefix caching disabled, the
     # processor builds overrides from request id instead of using user UUIDs.
-    processor = _mk_processor(monkeypatch, mm_cache_gb=0.0, enable_prefix_caching=False)
+    input_processor = _mock_input_processor(
+        monkeypatch, mm_cache_gb=0.0, enable_prefix_caching=False
+    )
 
     captured: dict[str, object] = {}
 
@@ -174,7 +175,7 @@ def test_multi_modal_uuids_ignored_when_caching_disabled(monkeypatch):
         return {"type": "token", "prompt_token_ids": [1]}
 
     monkeypatch.setattr(
-        processor.input_preprocessor, "preprocess", fake_preprocess, raising=True
+        input_processor.input_preprocessor, "preprocess", fake_preprocess, raising=True
     )
 
     request_id = "req-42"
@@ -188,7 +189,7 @@ def test_multi_modal_uuids_ignored_when_caching_disabled(monkeypatch):
         "multi_modal_uuids": mm_uuids,
     }
 
-    processor.process_inputs(
+    input_processor.process_inputs(
         request_id=request_id,
         prompt=prompt,  # type: ignore[arg-type]
         params=SamplingParams(),

vllm/engine/protocol.py

@@ -15,7 +15,7 @@ from vllm.sampling_params import SamplingParams
 from vllm.tasks import SupportedTask
 from vllm.transformers_utils.tokenizer import AnyTokenizer
 from vllm.v1.engine import EngineCoreRequest
-from vllm.v1.engine.processor import Processor
+from vllm.v1.engine.input_processor import InputProcessor
 
 
 class EngineClient(ABC):
@@ -23,7 +23,7 @@ class EngineClient(ABC):
 
     vllm_config: VllmConfig
     model_config: ModelConfig
-    processor: Processor
+    input_processor: InputProcessor
     io_processor: IOProcessor | None
 
     @property

vllm/entrypoints/llm.py

@@ -347,7 +347,7 @@ class LLM:
         self.supported_tasks = supported_tasks
 
         self.model_config = self.llm_engine.model_config
-        self.processor = self.llm_engine.processor
+        self.input_processor = self.llm_engine.input_processor
         self.io_processor = self.llm_engine.io_processor
 
     def get_tokenizer(self) -> AnyTokenizer:
@@ -364,7 +364,7 @@ class LLM:
             self.llm_engine.tokenizer = get_cached_tokenizer(tokenizer)
 
     def reset_mm_cache(self) -> None:
-        self.processor.clear_mm_cache()
+        self.input_processor.clear_mm_cache()
         self.llm_engine.reset_mm_cache()
 
     def get_default_sampling_params(self) -> SamplingParams:
@@ -1674,7 +1674,7 @@ class LLM:
             tokenization_kwargs,
         )
 
-        engine_request = self.processor.process_inputs(
+        engine_request = self.input_processor.process_inputs(
             request_id,
             engine_prompt,
             params,

vllm/entrypoints/openai/serving_engine.py

@@ -284,7 +284,7 @@ class OpenAIServing:
         self._async_tokenizer_pool: dict[AnyTokenizer, AsyncMicrobatchTokenizer] = {}
         self.log_error_stack = log_error_stack
 
-        self.processor = self.models.processor
+        self.input_processor = self.models.input_processor
         self.io_processor = self.models.io_processor
         self.model_config = self.models.model_config
         self.max_model_len = self.model_config.max_model_len
@@ -330,7 +330,7 @@ class OpenAIServing:
         return parser
 
     async def reset_mm_cache(self) -> None:
-        self.processor.clear_mm_cache()
+        self.input_processor.clear_mm_cache()
         await self.engine_client.reset_mm_cache()
 
     async def beam_search(
@@ -348,8 +348,8 @@ class OpenAIServing:
         length_penalty = params.length_penalty
         include_stop_str_in_output = params.include_stop_str_in_output
 
-        processor = self.processor
-        tokenizer = processor.tokenizer
+        input_processor = self.input_processor
+        tokenizer = input_processor.tokenizer
         if tokenizer is None:
             raise ValueError(
                 "You cannot use beam search when `skip_tokenizer_init` is True"
@@ -1214,7 +1214,7 @@ class OpenAIServing:
             self.max_model_len, params.truncate_prompt_tokens, tokenization_kwargs
         )
 
-        engine_request = self.processor.process_inputs(
+        engine_request = self.input_processor.process_inputs(
             request_id,
             engine_prompt,
             params,

vllm/entrypoints/openai/serving_models.py

@@ -69,7 +69,7 @@ class OpenAIServingModels:
             )
         self.lora_resolver_lock: dict[str, Lock] = defaultdict(Lock)
 
-        self.processor = self.engine_client.processor
+        self.input_processor = self.engine_client.input_processor
         self.io_processor = self.engine_client.io_processor
         self.model_config = self.engine_client.model_config
         self.max_model_len = self.model_config.max_model_len

vllm/v1/engine/async_llm.py

@@ -10,6 +10,7 @@ from typing import Any, cast
 
 import numpy as np
 import torch
+from typing_extensions import deprecated
 
 import vllm.envs as envs
 from vllm.config import VllmConfig
@@ -35,9 +36,9 @@ from vllm.utils.math_utils import cdiv
 from vllm.v1.engine import EngineCoreRequest
 from vllm.v1.engine.core_client import EngineCoreClient
 from vllm.v1.engine.exceptions import EngineDeadError, EngineGenerateError
+from vllm.v1.engine.input_processor import InputProcessor
 from vllm.v1.engine.output_processor import OutputProcessor, RequestOutputCollector
 from vllm.v1.engine.parallel_sampling import ParentRequest
-from vllm.v1.engine.processor import Processor
 from vllm.v1.executor import Executor
 from vllm.v1.metrics.loggers import (
     StatLoggerFactory,
@@ -112,7 +113,7 @@ class AsyncLLM(EngineClient):
         else:
             tokenizer = init_tokenizer_from_configs(self.model_config)
 
-        self.processor = Processor(self.vllm_config, tokenizer)
+        self.input_processor = InputProcessor(self.vllm_config, tokenizer)
         self.io_processor = get_io_processor(
             self.vllm_config,
             self.model_config.io_processor_plugin,
@@ -193,6 +194,14 @@ class AsyncLLM(EngineClient):
         else:
             self.profiler = None
 
+    @property
+    @deprecated(
+        "`AsyncLLM.processor` has been renamed to `AsyncLLM.input_processor`. "
+        "The old name will be removed in v0.13."
+    )
+    def processor(self):
+        return self.input_processor
+
     @classmethod
     def from_vllm_config(
         cls,
@@ -293,11 +302,7 @@ class AsyncLLM(EngineClient):
             request = prompt
         else:
             assert prompt_text is None
-            logger.warning_once(
-                "Processor has been moved under OpenAIServing and will "
-                "be removed from AsyncLLM in v0.13."
-            )
-            request = self.processor.process_inputs(
+            request = self.input_processor.process_inputs(
                 request_id,
                 prompt,
                 params,
@@ -481,7 +486,7 @@ class AsyncLLM(EngineClient):
         output_processor = self.output_processor
         log_stats = self.log_stats
         logger_manager = self.logger_manager
-        processor = self.processor
+        input_processor = self.input_processor
 
         async def output_handler():
             try:
@@ -532,7 +537,7 @@ class AsyncLLM(EngineClient):
                             engine_idx=outputs.engine_index,
                             scheduler_stats=outputs.scheduler_stats,
                             iteration_stats=iteration_stats,
-                            mm_cache_stats=processor.stat_mm_cache(),
+                            mm_cache_stats=input_processor.stat_mm_cache(),
                         )
             except Exception as e:
                 logger.exception("AsyncLLM output_handler failed.")
@@ -699,11 +704,11 @@ class AsyncLLM(EngineClient):
 
     @property
     def tokenizer(self) -> AnyTokenizer | None:
-        return self.processor.tokenizer
+        return self.input_processor.tokenizer
 
     @tokenizer.setter
     def tokenizer(self, tokenizer: AnyTokenizer | None) -> None:
-        self.processor.tokenizer = tokenizer
+        self.input_processor.tokenizer = tokenizer
 
     async def get_tokenizer(self) -> AnyTokenizer:
         if self.tokenizer is None:
@@ -738,7 +743,7 @@ class AsyncLLM(EngineClient):
         await asyncio.gather(*coros)
 
     async def reset_mm_cache(self) -> None:
-        self.processor.clear_mm_cache()
+        self.input_processor.clear_mm_cache()
         await self.engine_core.reset_mm_cache_async()
 
     async def reset_prefix_cache(self) -> None:

vllm/v1/engine/llm_engine.py

@@ -7,7 +7,7 @@ from copy import copy
 from typing import Any, cast
 
 import torch.nn as nn
-from typing_extensions import TypeVar
+from typing_extensions import TypeVar, deprecated
 
 import vllm.envs as envs
 from vllm.config import ParallelConfig, VllmConfig
@@ -28,9 +28,9 @@ from vllm.transformers_utils.tokenizer import AnyTokenizer, init_tokenizer_from_
 from vllm.usage.usage_lib import UsageContext
 from vllm.v1.engine import EngineCoreRequest
 from vllm.v1.engine.core_client import EngineCoreClient
+from vllm.v1.engine.input_processor import InputProcessor
 from vllm.v1.engine.output_processor import OutputProcessor
 from vllm.v1.engine.parallel_sampling import ParentRequest
-from vllm.v1.engine.processor import Processor
 from vllm.v1.executor import Executor
 from vllm.v1.metrics.loggers import StatLoggerFactory, StatLoggerManager
 from vllm.v1.metrics.reader import Metric, get_metrics_snapshot
@@ -88,7 +88,7 @@ class LLMEngine:
         else:
             tokenizer = init_tokenizer_from_configs(self.model_config)
 
-        self.processor = Processor(self.vllm_config, tokenizer)
+        self.input_processor = InputProcessor(self.vllm_config, tokenizer)
         self.io_processor = get_io_processor(
             self.vllm_config,
             self.model_config.io_processor_plugin,
@@ -135,6 +135,14 @@ class LLMEngine:
         # Don't keep the dummy data in memory
         self.reset_mm_cache()
 
+    @property
+    @deprecated(
+        "`LLMEngine.processor` has been renamed to `LLMEngine.input_processor`. "
+        "The old name will be removed in v0.13."
+    )
+    def processor(self):
+        return self.input_processor
+
     @classmethod
     def from_vllm_config(
         cls,
@@ -231,11 +239,7 @@ class LLMEngine:
             request = prompt
         else:
             assert prompt_text is None
-            logger.warning_once(
-                "Processor has been moved under LLM and will "
-                "be removed from LLMEngine in v0.13."
-            )
-            request = self.processor.process_inputs(
+            request = self.input_processor.process_inputs(
                 request_id,
                 prompt,
                 params,
@@ -307,7 +311,7 @@ class LLMEngine:
                 self.logger_manager.record(
                     scheduler_stats=outputs.scheduler_stats,
                     iteration_stats=iteration_stats,
-                    mm_cache_stats=self.processor.stat_mm_cache(),
+                    mm_cache_stats=self.input_processor.stat_mm_cache(),
                 )
                 self.do_log_stats_with_interval()
 
@@ -320,7 +324,7 @@ class LLMEngine:
         self.engine_core.profile(False)
 
     def reset_mm_cache(self):
-        self.processor.clear_mm_cache()
+        self.input_processor.clear_mm_cache()
         self.engine_core.reset_mm_cache()
 
     def reset_prefix_cache(self):
@@ -347,11 +351,11 @@ class LLMEngine:
 
     @property
     def tokenizer(self) -> AnyTokenizer | None:
-        return self.processor.tokenizer
+        return self.input_processor.tokenizer
 
     @tokenizer.setter
     def tokenizer(self, tokenizer: AnyTokenizer | None) -> None:
-        self.processor.tokenizer = tokenizer
+        self.input_processor.tokenizer = tokenizer
 
     def get_tokenizer(self) -> AnyTokenizer:
         if self.tokenizer is None: